File size: 2,653 Bytes
4553686
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
---
license: mit
tags:
- RAW
- RGB
- ISP
- NTIRE
- '2025'
- image
- processing
- low-level
- vision
- cameras
pretty_name: RAW Image Restoration Dataset
size_categories:
- 100M<n<1B
---

# RAW Image Restoration Dataset
## [NTIRE 2025 RAW Image Restoration](https://codalab.lisn.upsaclay.fr/competitions/21647)

- Link to the challenge: https://codalab.lisn.upsaclay.fr/competitions/21647
- Link to the workshop: https://www.cvlai.net/ntire/2025/

This dataset includes images **different smartphones**: iPhoneX, SamsungS9, Samsung21, Google Pixel 7-9, Oppo vivo x90. You can use it for many tasks, these are some:
- Reconstruct RAW images from the sRGB counterpart
- Learn an ISP to process the RAW images into the sRGB (emulating the phone ISP)
- Add noise to the RAW images and train a denoiser
- Many more things :)

### How are the RAW images?

- All the RAW images in this dataset have been standarized to follow a Bayer Pattern **RGGB**, and already white-black level corrected. 
- Each RAW image was split into several crops of size `512x512x4`(`1024x1024x3` for the corresponding RGBs). You see the filename `{raw_id}_{patch_number}.npy`.
- For each RAW image, you can find the associated metadata `{raw_id}.pkl`.
- RGB images are the corresponding captures from the phone i.e., the phone imaging pipeline (ISP) output. The images are saved as lossless PNG 8bits.
- Scenes include indoor/outdoor, day/night, different ISO levels, different shutter speed levels.

### How to use this?

- RAW images are saved using the following code:
  ```
  import numpy as np
  max_val = 2**12 -1
  raw = (raw * max_val).astype(np.uint16)
  np.save(os.path.join(SAVE_PATH, f"raw.npy"), raw_patch)
  ```
  We save the images as `uint16` to preserve as much as precision as possible, while maintaining the filesize small.

- Therefore, you can load the RAW images in your Dataset class, and feed them into the model as follows:
  ```
  import numpy as np
  raw = np.load("iphone-x-part2/0_3.npy")
  max_val = 2**12 -1
  raw = (raw / max_val).astype(np.float32)
  ```

- The associated metadata can be loaded using:
  ```
  import pickle
  with open("metadata.pkl", "rb") as f:
    meta_loaded = pickle.load(f)

  print (meta_loaded)
  ```

### Citation

Toward Efficient Deep Blind Raw Image Restoration, ICIP 2024

```
@inproceedings{conde2024toward,
  title={Toward Efficient Deep Blind Raw Image Restoration},
  author={Conde, Marcos V and Vasluianu, Florin and Timofte, Radu},
  booktitle={2024 IEEE International Conference on Image Processing (ICIP)},
  pages={1725--1731},
  year={2024},
  organization={IEEE}
}
```

Contact: [email protected]